How do you prevent architecture drift in long ChatGPT projects?
retained text
I keep hitting the same failure mode in long project chats: **fixed architecture → new plausible idea → model integrates it → a few turns later the new framing is treated as established** Nobody explicitly decided to change the design. The model just “helpfully” drifts. My current workaround is to prepend a short self-contained **architecture anchor** to every architecture-related message. It contains: \- current fixed design \- open decisions \- forbidden redesigns \- required next-step order \- a drift gate Core rule: *Before changing the architecture, identify the new evidence, real failure, or explicit user decision that justifies it. State what changes and what stays fixed. Otherwise:* ***STOP - ARCHITECTURE DRIFT.*** New chat ideas do not change project state by themselves. A small adversarial test worked: I suggested a “temporary assumption just for one dev test,” and ChatGPT correctly noticed that this would silently operationalize an unresolved design decision. Has anyone found a more robust way to handle this over long chats? External state files? Separate architect/reviewer chats? Periodic re-grounding? Something else?
document fetched 2026-09-27 00:29 · canonical https://reddit.com/r/OpenAI/comments/1wqck7m/how_do_you_prevent_architecture_drift_in_long